Why do Axon installation instructions refer to installing GoGi?

[Stephen Read]

Since the current version of axon seems to use cogentcore, why do the installation instructions refer to installing GoGi?

 

 

Stephen J. Read

Mendel B. Silberberg Professor of Social Psychology

Department of Psychology

University of Southern California

Los Angeles, CA 90089-1061

Website: www.stephenjread.com

 

 

 

[Randall O'Reilly]

Steve,

Thanks for pointing that out -- I just updated it.

Meanwhile, we are racing to get Leabra v2 working in time for your class in the fall, including updating the sims. We have a version of examples/ra25 in the v2 branch that is now fully working, along with major updates to the base emergent infrastructure to use the updated coding principles developed for the cogentcore framework, which cleans up a lot of things.

The new Leabra models leverage all of the "modern" emergent infrastructure in `looper` and `elog` etc, which manages the details of running over the different time scales and logging. This code has all been developed under axon, and makes the individual sim code cleaner, while still giving lots of flexibility for different use-cases.

The most exciting change will be that the sims can be run directly through a web browser via WebGPU (currently just Chrome, but other browsers will be introducing support soon). We are nearly done switching from Vulkan to WebGPU, and later will have the GPU compute functionality moved over to WebGPU too. WebGPU also as a "native" mode that runs on the desktop directly, and it is overall very similar to Vulkan (and runs over Vulkan on Linux), but just a lot simpler and cleaner. It also runs directly on Metal on a Mac, instead of requiring the annoying MoltenVK intermediate mapping layer. Installation on a Mac is now automatic! It appears just as performant in graphics, and I expect the same for compute.

Thus, soon, axon models will be runnable with GPU acceleration directly through the browser! Otherwise, the compute performance is significantly slower on web, because it runs through the web assembly (WASM) framework (and soon through Go -> javascript transpiling, which may be marginally faster). I do not plan to implement GPU compute support for Leabra, and at some point we will end up converting the textbook sims over to axon instead, which is really the future of this whole framework. Also, the Leabra v2 will not include PVLV for sure and perhaps other special algorithms -- these are already much improved in axon and it doesn't make sense to update in Leabra.

- Randy